I've running in the same problem. I'm using php-5.3.6 with opensuse11.3.

<?php
echo similar_text("test","wert");
echo "|";
echo similar_text("wert","test");

Expected result:
2|2
Current result:
1|2

Sounds like a bug, not a feature request, to me.

At first, it seems to be important to explain the algorithm. It is
finding the longest common substring of two strings, and then doing
this for the prefixes resp. the suffixes, recursively. The lengths
of the found common substrings are added and returned as result.
The percentage is calculated by multiplying the result with 200 and
dividing by the sum of the lenghts of both input strings.

Finding the longest common substring of two strings is accomplished
by the following algorithm:

    function longest_common_substring($a, $b) {
        $max = 0;
        for ($i = 0; $i < strlen($a); $i++) {
            for ($j = 0; $j < strlen($b); $j++) {
                $l = 0;
                while ($i + $l < strlen($a)
                       && $j + $l < strlen($b)
                       && $a[$i + $l] == $b[$j + $l]
                ) {
                    $l++;
                }
                if ($l > $max) {
                    $max = $l;
                }
            }
        }
        return $max;
    }

A noteworthy detail of the algorithm is that it always takes the
first longest common substring in either of the strings.

A simple example:
  similar_text('aabccc', 'aadccc') // => 5
  
The longest common substring is 'ccc' with a length of 3. None of
the input strings has a suffix after 'ccc', so only the prefixes
('aab' and 'aad') are compared. Their longest common substring is
'aa' (length 2); both don't have a prefix, so the suffixes 'b' and
'd' are compared, resulting in no common substring (length 0).
Adding up the lengths results in 5.

Another example:
  similar_text('michaelski', 'pankanin') // => 1
  
The first longest common substring that is found is 'i' (length 1),
so the prefixes are 'm' and 'pankan' and the suffixes are
'chaelski' and 'n'. The longest common substring of the prefixes is
'' (length 0), the longest common substring of the suffixes is ''
(length 0). So the result is 1.

Swapping the arguments has a different result, though:
  similar_text('pankanin', 'michaelski') // => 3
  
That is because the first longest common substring is 'a' (length
1), with the prefixes 'p' and 'mich' and the suffixes 'nkanin' and
'elski'. The prefixes have no common substring, but the suffixes
have 'k' (length 1), which has the prefixes 'n' and 'els' and the
suffixes 'anin' and 'i'. The prefixes have no common substring, but
the suffixes have 'i' (length 1). Therefore the result is 1.

One might argue that the algorithm is imperfect, but at least the
implementation is correct. It seems reasonable to document the
behavior of similar_text more clearly, therefore I'm changing to
"Documentation Problem".

> I used PHP 4.4.0 and 5.0.5 and got exactly the same results on
> both of them.

According to <http://3v4l.org/N6j98> the results under PHP 4.4.0
and 5.0.5 are the same than under most recent versions.

> Using LCS function (taken from comments on similar_text manual
> page)

This function does basically the same as longest_common_substring()
above, which is only part of the more complex similar_text()
algorithm.